Z.ai's GLM-5.3 reaches the agentic-coding frontier at a third of Kimi K3's size
Z.ai's GLM-5.3 passes Moonshot's Kimi K3 on many agentic-coding benchmarks — and Claude Fable 5 or GPT-5.6-Sol on some — with roughly 750B parameters, built by extending post-training on the GLM-5.2 base. Nathan Lambert credits release cadence over distillation: Z.ai ships in days where OpenAI and Anthropic spend months in pre-release testing. The weights reach Hugging Face in two weeks, putting near-frontier coding capability into open hands.
Source: interconnects.ai ↗
The time to release for Z.ai is likely days, not months as with OpenAI or Anthropic.
Nathan Lambert, interconnects.ai
Why this matters
- → Chinese labs now ship frontier coding capability in days, not months—undercutting American release cycles
- → Open weights reach Hugging Face in weeks, democratizing near-frontier agentic coding
- → Release velocity, not distillation or benchmaxxing, explains parity with vastly larger US models
Speed over scale